In early 2026 the Ministry of Health cut hospital computers off from external AI tools. The headlines called it a brake on progress. I read it as the opposite - a plain statement of the rules in the sectors where data is not allowed to wander. In healthcare and in finance the data will not travel to the AI, so the AI has to come to the data. The organizations that grasp this early build an advantage exactly where their competitors came to a stop.
Why the data really cannot leave
This is not caution. It is already a full legal framework. Amendment 13 to the Privacy Protection Law, in force since August 2025, carries fines of up to NIS 3 million and personal liability for officeholders in the event of a breach. The financial regulation anchored in December 2025 added a sharper principle: the supervised entity carries one hundred percent of the responsibility for whatever its AI system produces, even when the model belongs to an outside vendor. No surprise, then, that industry surveys heading into 2026 put privacy and governance at the top of the list of barriers to AI adoption in organizations (61.3% of respondents), well ahead of any concern about technical accuracy.
Now picture what that looks like on a Tuesday morning. A nurse pastes a patient summary into a free chatbot to draft a discharge letter. An analyst uploads a customer table into a cloud tool to tidy it up quickly. That information now sits on a foreign company's servers, sometimes feeds the training of its next model, and is permanently outside your control. Blocking the tools is a perfectly logical response to that scenario. The real management question is what you do the day after, because the need that drove the nurse to the chatbot has not gone away. It has gone underground, which is the more dangerous state of affairs.
A local model is not science fiction
An example from my own work rather than from a slide. I run a transcription setup that lives entirely on a local machine: a recorded conversation in Hebrew goes in, an accurate transcript with the speakers separated comes out, and not a single byte of audio reaches a cloud vendor. It runs on an ordinary processor, with no dedicated graphics card, and handles an hour of audio in under an hour. Marginal cost per conversation: zero. I built it because client recordings have no business leaving the machine. The economics surprise executives every time. A thousand hours of transcription a year through a paid cloud service is a recurring line item that grows with use; on hardware you own it is a small one-time investment, plus electricity.
What has changed is that open models - for transcription, for summarization, for pulling data out of documents - are now good enough for most operational work, and they run on hardware you control. In practice there are three levels of "inside the building": a model on a single workstation, for the most sensitive material; a dedicated server on the corporate network, for departmental use; and a private cloud with hard contractual commitments - a defined data region, zero retention, no training on your data. Good AI governance does not pick one level. It maps which kind of information is permitted at which level.
What you build inside the perimeter
Four families of tasks ripen first. None of them makes a clinical or financial decision on its own, and all of them give hours back.
Transcription and summarization: call center conversations, team meetings, committees. The audio stays inside and the minutes are ready the same day. Document quality control: comparing records and documentation against written procedures and flagging the gaps for a human to review. I have worked with control frameworks covering hundreds of procedural clauses, and this is work a model does with infinite patience and a person signs off on at the end. Mining the archive: old forms, scanned reports and letters turned into tables you can actually query. And preparing material for discussion: summarizing a long file ahead of a committee, with references back to the source.
One organizing rule: start where the AI saves hours and a human signs the output. That is how you pass the regulatory test and the organization's own trust test at the same time.
There is a further advantage that gets less airtime. Inside the perimeter you can finally give the AI the complete data. An external cloud model always receives a censored, thinned-out version of reality, because the part that matters is exactly the part you are not allowed to send it. A model inside the building sees the whole picture, which is why its output is worth using.
Five questions to ask before you sign with an AI vendor
If you sit on the management of a healthcare or financial organization, make every vendor answer these five in writing. Where does the data physically travel, and to which country? What does the vendor retain, and for how long? Is your data used to train models, its own or anyone else's? What happens the day you walk away: how do you get everything back, and what gets deleted? And who carries the liability when the model is wrong - a question the financial regulator has already answered on your behalf: you do.
A vendor who will not answer all five in writing does not come inside the perimeter. On the internal side, a short governance document setting out which kind of information is permitted at which level - local workstation, internal server, private cloud - saves you the same ad-hoc argument on every new project. And the pleasant surprise is how far you can get today without conceding a single one of the five.
Want to see a model running entirely inside your own network, on your own data, with nothing leaving the building? Send me a message - I would be glad to show you a demo.
